Researchers at the Lawrence Berkeley National Laboratory have created an artificial intelligence modeling framework that predicts how solid-state chemical reactions unfold over time. The system combines thermodynamic calculations with machine learning to model the full sequence of events in material synthesis, including intermediate compounds, final products, and unwanted impurities.
Published in Nature Materials, the research addresses a longstanding bottleneck in materials science. Synthesizing inorganic solid materials for batteries, sensors, and electronic devices currently requires mixing starting powders and heating them to temperatures as high as 800°C. This trial-and-error process can take anywhere from weeks to years because atoms move slowly through solid materials, often producing unintended compounds instead of the target product.
Accounting for kinetics
While existing computational tools evaluate thermodynamics to identify energetically stable materials, they struggle to make precise predictions because they omit kinetics—the speed and path of atomic movement through reaction sites. Kristin Persson, a senior scientist at Berkeley Lab and professor at the University of California, Berkeley, stated that atoms in solids travel much slower than in liquids, preventing reactions even when compounds are thermodynamically favored to form.
To overcome this limitation, the team developed a machine learning model trained to estimate atomic travel speeds across highly disordered solid interfaces. The model takes starting materials, mixing ratios, and temperature ramp-up schedules as inputs. Within minutes, it simulates the complete reaction pathway from beginning to end.
Testing on barium-titanium oxides
The researchers tested the system on barium-titanium oxides, a class of solid materials used in electronics. The team selected these compounds because their synthesis outcomes depend heavily on kinetics, as the new solid compounds formed during heating possess nearly identical thermodynamic stability.
The framework simulated reaction pathways across various starting ratios and temperature profiles. When compared against decades of published experimental synthesis data, the model matched the verified sequence of events, intermediate compounds, final products, and impurity formations.
Following the demonstration on barium-titanium oxides, the Berkeley Lab team plans to train and demonstrate the model on other classes of solid-state materials. The researchers eventually plan to build a foundation model using large kinetics datasets to support solid-state material discovery across multiple industries. The U.S. Department of Energy’s Office of Science funded the research.
